Abstract

We present a fully automatic approach to construct a large-scale, high- precision dataset from noisy web images. Within the entire pipeline, we focus on generating high quality seed images for subsequent dataset growing. High quality seeds are essential as we revealed, but they have received relatively less attention in previous works with respect to how to automatically generate them. In this work, we propose a density score based on rank-order distance to identify positive seed images. The basic idea is images relevant to a concept typically are tightly clustered, while the outliers are widely scattered. Through adaptive thresholding, we guarantee the selected seeds as numerous and accurate as possible. Starting with the high quality seeds, we grow a high quality dataset by dividing seeds and conducting iterative negative and positive mining. Our system can automatically collect thousands of images for one concept/class, with a precision rate of 95% or more. Comparisons with recent state-of-the-arts also demonstrate our method’s superior performance.